jingyaogong/minimind-v

Combines SigLIP2 vision encoding with lightweight MLP projection to fuse image features into a compact language model backbone, supporting both dense and mixture-of-experts architectures. Implements complete training pipeline including vision-language pretraining and supervised instruction-tuning with DDP multi-GPU acceleration, bfloat16 mixed precision, and checkpoint resumption across varying hardware configurations. Models range from 26M to 201M parameters with inference footprints under 1.1GB, packaged in Hugging Face Transformers format for ecosystem compatibility.

6,712 stars. Actively maintained with 16 commits in the last 30 days.